Computing with Nearby Mobile Devices A Work Sharing Algorithm for Mobile Edge-Clouds

Computing with Nearby Mobile Devices A Work Sharing Algorithm for Mobile Edge-Clouds

Abstract: As mobile devices evolve to be powerful and pervasive computing tools, their usage also continues to increase rapidly. However, mobile device users frequently experience problems when running intensive applications on the device itself, or offloading to remote clouds, due to resource shortage and connectivity issues.

Ironically, most users’ environments are saturated with devices with significant computational resources. This paper argues that nearby mobile devices can efficiently be utilized as a crowd-powered resource cloud to complement the remote clouds. Node heterogeneity, unknown worker capability, and dynamism are identified as essential challenges to be addressed when scheduling work among nearby mobile devices. We present a work sharing model, called Honeybee, using an adaptation of the well-known work stealing method to load balance independent jobs among heterogeneous mobile nodes, able to accommodate nodes randomly leaving and joining the system. The overall strategy of Honeybee is to focus on short-term goals, taking advantage of opportunities as they arise, based on the concepts of proactive workers and opportunistic delegator. We evaluate our model using a prototype framework built using Android and implement two applications.

We report speedups of up to 4 with seven devices and energy savings up to 71% with eight devices.

Existing system: There are several unique features that differentiate mobile crowd environments from a typical grid/distributed computing cluster, such as less computation power and limited energy on nodes, node mobility resulting in frequent disconnections, and node heterogeneity. Hence, solutions from grid/distributed computing cannot be used as they are, and need to be adapted to suit the requirements of mobile crowd environments. This paper presents the Honeybee model that supports P2P work sharing among dynamic mobile nodes. As proof of concept we present the Honeybee API, a programming framework for developing mobile crowd computing applications.

We build on previous work where we initially investigated static job farming among a heterogeneous group of mobile devices in , which was followed by a more self adaptive approach in using the ‘work stealing’ method , and in where three different mobile crowd sourcing applications were implemented and evaluated. The progress of our research on work sharing for mobile edge-clouds.

Proposed system: However, our results show that the work stealing method can provide better results. Social aware task farming has been proposed as an improvement on simple task farming, and social aware algorithms show better performance in their simulation based on real world human encounter traces. In the future we hope to build on this result (social aware task sharing) as an incentive for participation. In, human expertise is used to answer queries that prove too complicated for search engines and database systems, and in Crowd- Search, image search on mobile devices is performed with human validation via Amazon Mechanical Turk. A generic spatial crowd sourcing platform using smart phones is discussed in, where queries are based on location information. Mobile phones are used to collect sensor data on Medusa, according to sensing tasks specified by users. In Rankr, an online mobile service is used to ask users to rank ideas and photos. These are primarily concerned with the crowd sourcing aspect, using mobile devices as tools to access an online crowd sourcing service that is hosted on a remote server. In contrast,

Honeybee defines the crowd as the surrounding mobile devices and their users, and focuses on sharing the tasks on a crowd of local mobile devices with performance gain and saving energy as the main goal.

Advantages:

Each of the three methods have advantages depending mainly on the existence of high connectivity, additional infrastructure or node encounters respectively.

In our work, we focus on the third method, ie., opportunistically sharing work with the surrounding mobile devices, owing to issues with the other two approaches in cases of low network availability and lack of established infrastructure.

Furthermore, in Honeybee, we also recognize the potential of using mobile devices as agents of crowd sourcing, thereby exploiting the collective power of human expertise and machine resources.

Disadvantages: Unfortunately, these films do not have subtitles. Looking around his train, John realizes that the local passengers could help with his dilemma.

Collaboration among such smart mobile devices can pave the way for greater computing opportunities, not just by by creating crowd-sourced computing opportunities needing a human element, but also by solving the resource limitation problem inherent to mobile devices.

Modules: Mobile edge cloud: Today’s environments are becoming embedded with mobile devices with augmented capabilities, equipped with various sensors, wireless connectivity as well as limited computational resources. Whether we are on the move, on a train, or at an airport, in a shopping centre or on a bus, a plethora of mobile devices surround us every day], thus creating a resource-saturated ecosystem of machine and human intelligence. However, beyond some traditional web-based applications, current technology does not facilitate exploiting this resource rich

space of machine and human resources. Collaboration among such smart mobile devices can pave the way for greater computing opportunities, not just by by creating crowd-sourced computing opportunities needing a human element, but also by solving the resource limitation problem inherent to mobile devices. While there are research projects in areas such as mobile grid computing where mobile work sharing is centrally coordinate.

Crowd sourcing: This paper presents the Honeybee model that supports P2P work sharing among dynamic mobile nodes. As proof of concept we present the Honeybee API, a programming framework for developing mobile crowd computing applications.

We build on previous work where we initially investigated static job farming among a heterogeneous group of mobile devices in , which was followed by a more self adaptive approach in using the ‘work stealing’ method, and in where three different mobile crowd sourcing applications were implemented and evaluated. The progress of our research on work sharing for mobile edge-cloud.

With regards to the resource offloading, current research can be viewed from three main perspectives: offloading to a remote resource cloud, to a local cloudlet or local infrastructure, and to other mobile devices. Each of the three methods have advantages depending mainly on the existence of high connectivity, additional infrastructure or node encounters respectively. In our work, we focus on the third method, ie. Opportunistically sharing work with the surrounding mobile devices, owing to issues with the other two approaches in cases of low network availability and lack of established infrastructure.

Job scheduling method: The following characteristics of a mobile edge-cloud need to be considered when scheduling jobs among nodes: 1) heterogeneity: since nodes may be of heterogeneous capability and jobs may require varying amounts of resources, job allocation is non-trivial. Optimally stronger nodes should do more work. An expiration mechanismis needed so that stronger nodes can steal expired jobs taken by weaker nodes. Otherwise, if jobs were farmed equally, weak nodes may become bottlenecks. 2) Unknown capability: since the delegator is unaware of worker capability, it is not possible for the delegator to assign more work to stronger

nodes. Exchanging metadata is not effective due to node dynamism, e.g., the node capabilities may change randomly, thereby making the information derived from metadata invalid. 3) Dynamism: due to mobility and factors such as human intervention and low battery, nodes are prone to failure. Hence the possibility of frequently disconnections and new nodes randomly joining need to be supported, and the overall strategy needs to focus on short term goals and take advantage of opportunities as they arise.

Facial recognition at the workplace: Even a resource rich environment can sometimes be unusable depending on the type of work and the location of data. For example, take the case of Jane at her office equipped with many PCs. She has hundreds of photographs on her camera phone, taken at a recent office party. Her friend Mary asks to send her all the photographs where Mary appears. As Jane is unable to connect her phone to a computer, she considers using her the device to run a facial recognition app that compares each of the images with a photograph of Mary and filters the ones containing her face. However facial recognition algorithms are costly, and processing a large number of photographs could take a substantial amount of time, freeze the device and drain the battery. These challenges can be addressed by offloading/crowd sourcing the image processing task to external computing resources, as explored implemented in projects such as GeoTag-X6 and Galaxy Zoo7. However, the aforementioned approaches need remote clouds and/or Cloudlets, neither of which are available to Jane. Therefore Jane employs Honeybee to share the task with mobile devices belonging to her colleagues.

Worker heartbeat: The dynamic nature of mobile edge-clouds will incur frequent and unpredictable disconnections. If the disconnections are unidentified, the delegator may wait unnecessarily for the return of stolen jobs. To address this, each worker sends a periodic signal to indicate that it is alive. If the worker had sent results to the delegator, or acknowledged a job transmission, within that time period they are also counted as heartbeats. If the delegator cannot hear a worker heartbeat for m consecutive checks, the delegator deems that worker is dead, i.e., either moved away or lost connectivity, and adds the respective stolen jobs back to the queue.

Performance gain: Both applications were tested for speedups for a fixed task size while varying the number of devices. Figure 7a gives the face detection performance results for 960 images. As can be seen, the speedup is proportional to the number of devices and the maximum average speedup observed was 4:012 for 7 devices. Results from Mandelbrot set generation show a similar trend in Figure 7b where the maximum average speedup was 3.724. Figure 7c shows the speedups for Face detection using varying numbers of jobs versus a fixed number of devices. Comparing the results for both 2 and 3 devices, in both cases the speedup increases proportionally to the task size. Figures 7dand 7e show the percentage and amount of time saved for Face detection, and Figure 7f shows the percentage of time saved for Mandelbrot set generation. As can be seen from all three graphs, the time saved is proportional to task size (total number of jobs) and amount of resources. From Figures 7a, 7b, 7d and 7f, it is clear that the speedup plateaus after reaching the maximum speedup value.